Repository navigation
docs(proposal): add SchedulerProfile design proposal - #1926
Conversation
f660977 to
471fdaa
Compare
471fdaa to
5c24c75
Compare
848dd97 to
7934060
Compare
Signed-off-by: jasoncao <w_caoyu@kingsoft.com>
Signed-off-by: jasoncao <w_caoyu@kingsoft.com>
Signed-off-by: jasoncao <w_caoyu@kingsoft.com>
Signed-off-by: jasoncao <w_caoyu@kingsoft.com>
Signed-off-by: jasoncao <w_caoyu@kingsoft.com>
7934060 to
8655b81
Compare
…ht gating Signed-off-by: jasoncao <w_caoyu@kingsoft.com>
|
|
||
| Users migrating from the global config must copy every plugin and arg | ||
| they rely on into the profile — nothing carries over. This differs from | ||
| the semantics discussed for the field-based approach in #1899, where a |
There was a problem hiding this comment.
Don't need to mention these, only add to the confusion.
There was a problem hiding this comment.
Don't need to mention these, only add to the confusion.
Okay, I will modify it
There was a problem hiding this comment.
We could keep Users migrating from the global config must copy every plugin and arg they rely on into the profile — nothing carries over.
There was a problem hiding this comment.
We could keep
Users migrating from the global config must copy every plugin and arg they rely on into the profile — nothing carries over.
OK👌
| weight ranges. | ||
| - *Plugin-dependent*: plugin names must be registered, args must pass | ||
| each plugin's validator, plugins must support their declared role | ||
| (filter vs score), and the random plugin must be the only score |
There was a problem hiding this comment.
why the random plugin must be the only score plugin?
There was a problem hiding this comment.
why
the random plugin must be the only score plugin?
This is the existing logic, which is mentioned in the handleRandomPluginConflicts method of /pkg/kthena-router/scheduler/plugins/conf/conf.go.
There was a problem hiding this comment.
Okay, that's ambiguous.
I thought it meant the score plugin could only use random.
We can delete this.
There was a problem hiding this comment.
Okay, that's ambiguous.
I thought it meant the score plugin could only use random.
We can delete this.
Agreed, it read ambiguously. Removed from the checklist
|
two nits, generally lgtm |
Per review feedback. Signed-off-by: jasoncao <w_caoyu@kingsoft.com>
…otes Signed-off-by: jasoncao <w_caoyu@kingsoft.com>
|
/lgtm @0xjasoncao Thanks for your contribution :) |
|
[APPROVALNOTIFIER] This PR is APPROVED This pull-request has been approved by: YaoZengzeng The full list of commands accepted by this bot can be found here. The pull request process is described here DetailsNeeds approval from an approver in each of these files:
Approvers can indicate their approval by writing |
What type of PR is this?
/kind documentation
What this PR does / why we need it:
Adds the design proposal for
SchedulerProfile— per-ModelServer schedulerplugin combinations. A SchedulerProfile holds a complete scheduler plugin
combination (filter and score plugins with weights and args); ModelServers
reference a profile by name, with fallback to the global scheduler config.
The design was discussed in #1899 and the 2026-10-27 community meeting;
direction confirmed by @YaoZengzeng. Implementation will follow in a
separate PR.
Which issue(s) this PR fixes:
Part of #1899
Bug evidence (required for bug-related PRs):
N/A
Special notes for your reviewer:
docs/proposal/), no code impact.Does this PR introduce a user-facing change?: